Cerebral Cortex, May 2018;28: 1771–1782

doi: 10.1093/cercor/bhx087
Advance Access Publication Date: 10 April 2017
Original Article

ORIGINAL ARTICLE

Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020

Bayesian Mapping Reveals That Attention Boosts Neural
Responses to Predicted and Unpredicted Stimuli
Marta I. Garrido1,2,3,4, Elise G. Rowe1,2, Veronika Halász1,2,3 and
Jason B. Mattingley1,3,5
1
Queensland Brain Institute, The University of Queensland, 4072 Brisbane, Australia, 2Centre for Advanced
Imaging, The University of Queensland, 4072 Brisbane, Australia, 3ARC Centre of Excellence for Integrative
Brain Function, The University of Queensland, 4072 Brisbane, Australia, 4School of Mathematics and Physics,
The University of Queensland, 4072 Brisbane, Australia and 5School of Psychology, The University of
Queensland, 4072 Brisbane, Australia
Address correspondence to Marta I. Garrido, Queensland Brain Institute, The University of Queensland, Building 79, St Lucia 4072, Brisbane, Australia.
Email: m.garrido@uq.edu.au

Abstract
Predictive coding posits that the human brain continually monitors the environment for regularities and detects
inconsistencies. It is unclear, however, what effect attention has on expectation processes, as there have been relatively few
studies and the results of these have yielded contradictory findings. Here, we employed Bayesian model comparison to
adjudicate between 2 alternative computational models. The “Opposition” model states that attention boosts neural
responses equally to predicted and unpredicted stimuli, whereas the “Interaction” model assumes that attentional boosting
of neural signals depends on the level of predictability. We designed a novel, audiospatial attention task that orthogonally
manipulated attention and prediction by playing oddball sequences in either the attended or unattended ear. We observed
sensory prediction error responses, with electroencephalography, across all attentional manipulations. Crucially, posterior
probability maps revealed that, overall, the Opposition model better explained scalp and source data, suggesting that
attention boosts responses to predicted and unpredicted stimuli equally. Furthermore, Dynamic Causal Modeling showed
that these Opposition effects were expressed in plastic changes within the mismatch negativity network. Our findings
provide empirical evidence for a computational model of the opposing interplay of attention and expectation in the brain.

Key words: EEG, MMN, modeling, novelty, prediction

Introduction attention is the process of prioritizing information by allocating

The way in which we perceive the world around us is thought to
be an active inferential process. Rather than passively registering
information that arrives at our senses, the brain builds predictive
models of what it might encounter next. These theoretical con-
jectures have been formalized in terms of predictive coding (Rao
and Ballard 1999; Friston 2005) and are useful in explaining the
ubiquitous phenomenon of larger brain responses to surprising
than predictable events (Montague 1999; Garrido et al. 2013)
(Opitz et al. 1999; Summerfield and Koechlin 2008). Selective
more cognitive resources to the object of focus, while suppres-
sing information that is irrelevant. Recent extensions of predic-
tive coding have framed attention as the process of enhancing
the reliability of prediction errors (Feldman and Friston 2010).
This idea has been empirically demonstrated by larger prediction
errors for attended than unattended visual objects (Jiang et al.
2013) and sounds (Auksztulewicz and Friston 2015), with the lat-
ter going against the longstanding notion of mismatch negativity
(MMN) as a preattentive process (Naatanen et al. 2001).

© The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com
 1772 | Cerebral Cortex, 2018, Vol. 28, No. 5
There is a general consensus that expectation dampens
neuronal activity and that attention boosts neuronal activity
(Summerfield and Koechlin 2008). Thus, superficially at least,
attention and prediction appear to have opposing effects.
However, the way in which attention interacts with expectation
is unclear for 2 reasons. First, there have been very few
attempts to manipulate attention and prediction indepen-
dently, but many instances in which the 2 have been entwined
or confounded (Summerfield and Egner 2009), as attention is
often manipulated in a probabilistic manner rather than
through stimulus filtering or prioritization. Second, the few
studies on prediction and attention have yielded a puzzling
depiction of what might be happening in the brain. Kok et al.
(2012) provided fMRI evidence that attention and prediction
have an interactive or synergetic effect by showing greater
brain activity in the visual cortex for predicted (than unpre-
dicted) visual stimuli, a finding which was conceptually repli-
cated using electroencephalography (EEG) for auditory stimuli,
and expressed in the N1 evoked potential (Hsu et al. 2014). By
contrast, (Auksztulewicz and Friston 2015) found that attention
increased the typically observed difference between evoked
responses to unpredicted versus predicted stimuli, as reflected
in an enhanced MMN, and Bekinschtein et al. (2009) found that
violation of global rules led to late evoked responses only when
participants were aware of such violations.
In this paper, we first formalize 2 theoretical models that
have been put forward to explain the interplay between atten-
tion and prediction in the brain: the Opposition model and the
Interaction model, introduced in Kok et al. (2012). The
Interaction model postulates that attention and prediction
interact such that neuronal activity is greatest for attended and
predicted events. This model is inspired by the idea that atten-
tion increases the precision of predictions by weighting predic-
tion errors (Feldman and Friston 2010), and assumes 4 levels of
precision, or attention, that depend on the level of prediction.
By contrast, the Opposition model posits that attention and
prediction have opposing effects on neural activity, such that
prediction mitigates and attention boosts neural activity. The
predictions of this model are that the neuronal responses will
be greatest for attended unpredicted stimuli, and smallest for
unattended predicted stimuli. Computationally, this model
assumes that neuronal activity is weighted by 2 (instead of 4)
levels of attention (attended and unattended). This model is
agnostic about the relationship between responses to
attended predicted and unattended unpredicted events. Both
the Interaction and the Opposition models assume that pre-
diction has 2 levels, such that unpredicted stimuli evoke a
larger neuronal response than predicted stimuli. They differ,
however, in their treatment of the attention component.
Specifically, the Opposition model offers a more parsimonious
expression of the effects of attention on neuronal responses
(Fig. 1).
Here we tested these models empirically using Bayesian
model comparison for scalp and source EEG data, as well as
dynamic causal modeling (DCM). We developed a novel audi-
tory task in which participants were presented with indepen-
dent streams of white noise concurrently in each of the 2 ears,
and were instructed to attend to the left channel, the right
channel or both channels in separate blocks to detect brief gaps
in the noise streams. At the same time, an irrelevant stream of
standard and deviant tones was presented in either ear
(attended or ignored), providing an orthogonal stimulus set
from which to extract neural responses to predicted and unpre-
dicted auditory events.
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
Figure 1. Two competing models for the relationship between attention and
prediction. In the Opposition model, predicted (green) and unpredicted (red)
neural signals are multiplied by 2 levels of Attention, with attended stimuli
(solid lines) receiving a greater boost than unattended stimuli (dashed lines). In
the Interaction model (proposed by Kok et al. 2012), predicted and unpredicted
signals are multiplied by 4 (instead of 2) levels of attention that depend on the
level of the prediction.
Methods
Participants
A total of 21 healthy adults were recruited for the experiment.
Data from 2 participants were excluded from further analysis
due to poor performance on the behavioral task (accuracy
< 50%). The reported analysis was thus performed on data
from 19 participants (10 females, aged 19–43, M = 24.21, stan-
dard deviation = 6.11) with no reported history of neurological
or psychiatric disorder and no previous head trauma resulting
in unconsciousness. All participants gave written informed
consent in accordance with the guidelines of the University
of Queensland’s ethical committee, and were monetarily com-
pensated for their time.
Auditory Stimuli
The auditory task developed for the study is depicted in Figure 2.
An auditory frequency oddball sequence was played to one ear
at 60 dB and overlayed with Gaussian white noise at 40 dB. White
noise only was played to the other ear at 40 dB. Two pure tones,
standards (P = 0.85) and deviants (P = 0.15), (f = 500 or 550 Hz;
counter-balanced between blocks) of 50 ms in duration were
played with an inter-stimulus interval of 450 ms. Embedded in
the white noise of either ear were 2 types of targets: a total of 30
nonoverlapping randomized periods of no sound (gaps), which
could be singular (90 ms gaps only, 15 per block) or doubled (two
90 ms breaks separated by a 30 ms white noise return, 15 per
block). The gaps in the white noise of either ear were never
within 2.5 s of each other and never occurred at the same time as
a tone. Importantly, the presentation timings for the noise gaps
and the tones were uncorrelated in order to avoid any systematic
effects of bottom-up attention that could have otherwise con-
founded the ERPs to the tones. All auditory stimuli were created
using in-house Matlab scripts, recorded using Audacity Sound
Mixer prior to the experiment, and delivered with inner-ear buds
(Etymotic, ER3).
Experimental Design
The twelve experimental trial blocks (T = 3:32 min each) were
comprised of a total of 380 tones (with deviants always falling
within 4–10 standard tones). Participants were instructed to

Figure 2. Experimental paradigm. Gaussian white noise embedded with single
(90 ms) or double (210 ms) noise gaps (periods of silence, and the targets of this
experiment) was played to both ears (different target sequence in each ear).
One ear received the oddball sequence of pure tones (50 ms) at either 500 or
550 Hz (counter-balanced between blocks) (ISI = 450 ms, standard P = 0.85 black
rectangle, deviant P = 0.15 hollow rectangle, respectively). Participants were
instructed to pay attention to the targets embedded in the white noise in the
left, right, or both ears and to ignore the tones. ISI = inter-stimulus interval;
L = left ear; R = right ear; Std = standard; Dev = deviant.
listen for and report target gaps within the white noise stream
in either the left channel only, the right channel only, or in
either channel (divided attention), and to ignore the tones.
Each attention condition was repeated 4 times and the order
of the blocks was pseudo-randomized such that no partici-
pant received the same order. When a target was identified in
the attended ear(/s) participants responded with a “1” key-
press if the gap was singular and a “2” keypress if the gap was
doubled. In one-third of the blocks oddball tones were played
in the attended ear, in another third the tones were played in
the ignored ear, and in the remaining third, in which partici-
pants divided their attention between ears, the tones were
presented to either side, counter-balanced between the left
and right across separate blocks. Participants performed all
blocks in one testing session of 60 min (42:24 min total task
duration plus breaks) with an additional 30 min EEG setup
period.
Modeling Attention and Prediction Garrido et al. | 1773
Task
Participants were seated in front of a computer screen and
wore inner-ear buds for the duration of the experiment. Prior to
recordings, participants listened to an example auditory stream
of 1-min duration, which demonstrated the single and double
gaps in the white noise. Each participant then underwent a
brief practice session with auditory stimuli consisting of 9 sin-
gle and 9 double gaps, and a total of 110 tones. Participants
were given feedback about their accuracy in this practice block
but not in the experimental blocks. At the beginning of each
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
experimental block, the focus of attention was specified ver-
bally and an arrow (left, right, or both directions) remained on
the screen for the duration of the block as a reminder.
Participants were asked to make their keypresses in response
to target gaps as quickly and as accurately as possible, and to
ignore any gaps in the uncued ear (in the focused attention
condition). Task performance was assessed based on the per-
centage of correctly detected target gaps and reaction times.
Participants with <50% overall accuracy (proportion correct)
were excluded from further analysis.
EEG Data Acquisition and Preprocessing
Continuous EEG data were recorded with a Biosemi Active Two
system with 64 Ag/AgCl scalp electrodes arranged according to
the international 10-10 system for electrode placement using a
nylon head cap. Data were recorded at a sampling rate of
1024 Hz. Preprocessing and data analysis were performed with
SMP12 (http://www.fil.ion.ucl.ac.uk/spm/). Data were rereferenced
to a common reference, down-sampled to 200 Hz and high-pass
filtered above 0.5 Hz. Eye blinks were detected and marked using
the VEOG channel before the data were epoched offline with a
peristimulus window of −100 to 400 ms. Artefact removal was
performed by removing trials marked with an eyeblink and by
thresholding all channels at 100 uV. Trial data were robustly
averaged before being low-pass filtered below 40 Hz and base-
line corrected between −100 and 0 ms. We analysed event-
related potentials with respect to the onsets of standard and
oddball tones, separately for conditions in which the tones
were presented in the attended ear, the unattended ear, or in
either ear in the divided attention condition.
Spatiotemporal Image Conversion
Event-related potentials were converted into 3D spatiotemporal
volumes per condition and participant. This was achieved by
interpolating and dividing the scalp data per time point into a
2D 32 × 32 matrix. We obtained one 2D image for every time
bin (from 0 to 400 ms in steps of 5 ms). These images were then
stacked according to their peristimulus temporal order, result-
ing in a 3D spatiotemporal image volume with dimensions of
32 × 32 × 81 per participant. Data were then smoothed at
FWHM 12 × 12 × 20 mm3.
Spatiotemporal Statistical Maps
For each participant, the 3D spatiotemporal image volumes
were modeled with a mass univariate general linear model
(GLM) as implemented in SPM12. We performed between-
subject F-contrasts for (1) the main effect of attention, (2) the
main effect of prediction, and (3) the interaction between atten-
tion and prediction. Simple effects were estimated using
between-subject t-statistic contrasts. The same statistical anal-
yses were performed on the 3D spatial image volume obtained
 1774 | Cerebral Cortex, 2018, Vol. 28, No. 5
after source localization (see below). All sensor effects are
reported at a threshold of P < 0.05 with family-wise error (FWE)
correction for multiple comparisons over the whole spatiotem-
poral volume. For closer inspection of the main effects and
interactions obtained at channel Fz (at which predictability
effects are typically strongest, Naatanen and Alho (1997)), we
implemented a 1D GLM approach using SPM12. We restricted
our time window from 0 to 400 ms after stimulus onset and, in
a separate analysis, between the typical MMN time window of
100–250 ms (FWE corrected over the time bins considered).
Source Reconstruction
We obtained source estimates on the cortical mesh by recon-
structing scalp activity with a Boundary Element Method (BEM)
and a standard MNI template for the cortical mesh, in the
absence of individual MRIs. This forward model was then
inverted with multiple sparse priors (MSP) assumptions for the
variance components under group constraints. This allowed for
inferences on the most likely cortical regions that generated
the sensor-level data. We obtained images from these recon-
structions for each of the 6 conditions in every participant.
These images were smoothed at FWHM 12 × 12 × 12 mm3. We
then computed the main effects of attention and prediction,
and the interaction (attention × prediction) using conventional
SPM analysis. The effect of prediction (t-statistic) is displayed at
an uncorrected threshold of P < 0.001. These weaker signifi-
cance criteria were used for post-hoc visualization, once the
effects had been established under robust criteria at the scalp
level, and we only report regions significant at P < 0.05 FWE cor-
rected at the cluster level.
Statistics
Significance sensor space maps for prediction effects are dis-
played at P < 0.05 corrected for multiple comparisons using
family-wise error rate. The interaction map is displayed at P <
0.01 uncorrected for the purpose of defining a region of interest
for follow up Bayesian Model Selection (BMS). Source maps are
displayed at P < 0.001 uncorrected, but only significant cluster-
level PFWE < 0.05 are reported.
BMS was employed to make inferences on both scalp and
source maps, as well as on DCMs. Note that this framework
uses model evidence as a relative (probabilistic) measure for
how well one model explains the data relative to another, con-
sidered in the model space. Importantly, model evidence seeks
the optimal balance between accuracy and model complexity,
by favouring the former and penalizing the latter.
Bayesian Model Selection
To compare the 2 models (Opposition and Interaction; see
Introduction) of the effects of attention on prediction (standard
and deviant tones) we used the BMS methodology described in
Rosa et al. (2010), and adapted here for EEG. For this analysis
we discarded trials from the divided attention condition and
used only the attended and unattended trials from the focused
attention conditions (attend left ear only, attend right ear only)
for both standard and deviant tones. We created posterior
probability maps (PPMs) from individual participant log-model
evidences using a random-effects approach (RFX). Here, the
winning model was the one with the highest log-evidence
(assuming uniform priors over the models) across participants.
We performed this analysis at the sensor and source levels by
modeling the data with regressors describing the hypothesized
relationships amongst the 4 different conditions.
Briefly, covariate regressor weights were applied to every par-
ticipant and trial under the Opposition model, which predicts
reductions in ERP amplitudes across conditions in the following
order: (1) attended unpredicted, (2) unattended unpredicted/
attended predicted, and (3) unattended predicted. Next, we speci-
fied a second model derived from Kok et al. (2012), the Interaction
model, which predicts reductions in ERP amplitudes across condi-
tions in the following order: (1) attended predicted, (2) attended
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
unpredicted, (3) unattended unpredicted, and (4) unattended pre-
dicted. Voxel-wise whole-brain log-model evidence maps were
then created for every participant and model, estimated using the
Variational Bayes first-Level Model Specification methodology
described in Penny et al. (2005). Source level maps were further
smoothed with a 1 mm half width Gaussian kernel. We used the
RFX approach to produce PPMs for both models at the group-
level. These maps (displayed at a threshold of probability larger
than 75% and 50% for scalp and source, respectively) allowed us
to compare which model had the higher probability at each voxel
in the brain (and at each time point in the scalp level analysis).
Further model comparisons for specific regions at the sensor level
were undertaken using brain regions selected a priori from the
attention by prediction interaction contrast. At the source, these
comparisons were made at the peak coordinates of clusters for
each model that exceeded 51%.
Dynamic Causal Modeling
Source locations were identified based on multiple sparse
priors source reconstruction of the overall mismatch (P < 0.05
uncorrected threshold). These regions were: bilateral primary
auditory cortices (A1; MNI coordinates: left [−42, −24, 34] and
right [44, −22, 38]), bilateral inferior temporal gyri (ITG; MNI coor-
dinates: left [−42, −10, −38] and right [44, 0, −42]) and left inferior
frontal gyrus (LIFG; MNI coordinates: [−50, 32, 0]). The choice for
the DCM source nodes was undertaken in a data-driven fashion,
rather than being based on a priori models drawn from the litera-
ture. Previous papers have consistently tested models with A1,
STG, and IFG, nodes that we first proposed in 2007 based on a
number of fMRI and MEG studies (Garrido et al. 2007). The simi-
larity of the model space proposed here and in previous papers is
evident, except for the replacement of STG with ITG. This was
motivated by the strong evidence for ITG both in the standard
SPM analysis at the source level (P < 0.05, FWE-cluster corrected),
and the posterior probability maps (probability >80%). It is impor-
tant to note, however, that while some papers have tested mod-
els pertaining to the presence or absence of nodes such as STG
and IFG, all assumed that these nodes were correct, given pre-
vious literature, rather than refining the model progressively
through exploring other candidate nodes (e.g., ITG instead of
STG) through DCM optimization or other forms of source recon-
struction. Given the strong evidence from our source recon-
structed data and the posterior probability maps, as well as the
novelty of the paradigm, here we took a data-driven approach
rather than adhering to an assumed model specification.
Nevertheless, we ran a validation check to rule out the possibil-
ity that our source reconstructed nodes were less reliable than
the nodes taken a priori from the literature. BMS revealed that
the models with the source reconstructed nodes outperformed
the models with a priori nodes by 60% probability.
Note that in the absence of individual anatomical land-
marks, we used a standard MNI template for the cortical mesh
in our source reconstruction, which was then used to identify

candidate nodes for the DCMs. Whether including anatomical
information would improve the source reconstruction results
at the group level is unclear. This raises an interesting model
comparison related to that addressed in Mattout et al. (2007);
Henson et al. (2009), who showed that individual MRI does not
add to the precision of source estimates compared with an
individual deformed template. This was done for MEG data,
however, and it is unclear what the impact on EEG might be
when using an MNI template without individual deformations.
Given that MEG has higher spatial resolution and is more sensi-
tive to approximations in source models, however, it is likely
that any potential benefit afforded by individual MRIs would be
smaller (not larger) for EEG than for MEG (shown to be negligi-
ble). Furthermore, the sensitivity of our group level inference
yielded a reconstruction of the expected brain regions underly-
ing the MMN (within the temporal and inferior frontal cortex),
even in the absence of a highly realistic head model. Importantly,
the locations of the source reconstruction were only used as soft
priors in the subsequent DCM analysis, so that source locations
could be adjusted individually during the connectivity estimation
procedure.
We first optimized the basic connectivity architecture using
responses to attended and unattended standards and deviants
with no between-trial effects present. This first step considered
2 competing model structures that included bilateral A1 and
ITG, but differed in the presence or absence of LIFG. Next, the
pattern of changes in extrinsic connectivity was optimized
under the fully connected architecture (the winning model)
using responses in all 4 conditions for the Opposition and
Interaction models. The family of Opposition models used a
between-trial effect of [1, 2, 2, 3] for the attended predicted, pre-
dicted attended, unpredicted unattended, and attended unpre-
dicted, respectively. The family of Interaction models, on the
other hand used [1, 2, 3, 4] for predicted unattended, unpredicted
unattended, unpredicted attended, and attended predicted,
respectively. The choice of the weights for the between-trial
effects was motivated by the theoretical relationship proposed
in Figure 1. It is important to note that there is an infinite num-
ber of possible combinations of weights that could satisfy the
general ordinal relationship between the 4 conditions specified
in Figure 1. Here, we assumed a linear relationship, in the
absence of theoretical or empirical evidence to assume an other-
wise more complex relationship. Specifically, for the Opposition
Model, we compared models using [1 2 2 3] or [1 3 3 4], as we had
no reason to believe that the attended predicted and the unat-
tended unpredicted conditions would be closer to the unat-
tended predicted condition than to the attended unpredicted
condition. Bayesian model comparison revealed that the former
outperformed the latter.
Fifteen competing models were tested, each with a different
subset of connections—forward (F), backward (B), and recurrent
(R)—which also included (subscript i) or excluded intrinsic modu-
lations of A1, and a single null model. Finally, the Opposition and
Interaction model-dependent changes in intrinsic connectivity
were then grouped by families, under the optimized connectivity
architecture. In both DCM estimation steps, models were
inverted using a 0–400 ms peristimulus time window.
Results
Behavioral Findings on Attentional Manipulation
Behavioral results for the target detection task—discriminating
single- and double-gaps in concurrent white noise streams in
Modeling Attention and Prediction Garrido et al. | 1775
each ear—were grouped into unilateral (focused) or bilateral
(divided) attention conditions (30 targets over 8 blocks and 60
targets over 4 blocks, respectively). We excluded any partici-
pants who did not achieve mean response accuracy >50%.
There was no significant difference in response accuracy (P =
0.14) between the unilateral (M = 71.80%, standard error of
mean [SEM] = 5.19%) and the bilateral (M = 68.33%, SEM =
5.13%) conditions. Participants were significantly faster (P =
0.03) to respond in the bilateral (M = 748.16 ms, SEM = 27.67 ms)
than the unilateral conditions (M = 779.79 ms, SEM = 34.13 ms),
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
likely reflecting a strategy of responding immediately to any
target gap when monitoring both ears under divided attention,
as opposed to having to select only relevant gaps in the focused
attention conditions (filtering out gaps in the ignored ear).
Attention Amplifies Prediction Errors—Single
Channel Analysis
ERPs corresponding to each of the experimental conditions (as
well as the MMNs derived from subtracting the standards from
the deviants within a condition) were extracted from electrode
Fz and compared over time (Fig. 3). The N1 and P2 components
were examined as an average across participants and condi-
tions. For this, the lowest time point between 50 and 150 ms
and highest point between 150 and 250 ms were determined
from the omnibus ERP plot (i.e., the mean ERP across all partici-
pants and conditions over time). These time indexes ±25 ms
were then used to find the average ERP per condition. Statistical
tests of the N1 components found only a main effect of surprise
(F[1,72] = 4.9583, P = 0.0291). Similarly, at P2, there was a main
effect of surprise (F[1,72] = 17.5898, P = 7.7001e-05), but no fur-
ther significant main effects or interactions. In addition, results
at Fz from 0 to 400 ms using the 1D GLM approach revealed a
significant main effect of Attention between 290 and 340 ms
(PFWE_cluster = 0.006), and a significantly larger prediction error
for attended relative to unattended conditions at 115–120 ms
(PFWE_cluster = 0.020). We then restricted our analysis to the MMN
time window (100–250 ms) and again found a significant main
effect of Attention but at an earlier period between 200 and
230 ms (PFWE_cluster = 0.028). Moreover, there was a significantly
larger prediction error for attended versus unattended condi-
tions between 100 and 130 ms (PFWE_cluster = 0.046). These find-
ings demonstrate that attention amplifies prediction errors.
Larger Responses to Unpredicted Than Predicted Events
Regardless of Attention Level—Sensor and Source Space
As shown in Figure 4A, the main effect of Prediction, or surprise
(standards vs. deviants), disclosed several significant compo-
nents comprised of 2 late effects. The first late effect was
detected from 200 to 220 ms (peak-level Tmax = 8.30, cluster-
level PFWE < 0.001; at frontocentral channels). The second late
component was observed from 290 to 295 ms (peak-level Tmax =
5.02, cluster-level PFWE = 0.004; at right parieto-occipital chan-
nels). We also found simple Prediction effects in all of the
attention manipulations (Fig. 4B), that is, attended (peaking at
185 ms), unattended (peaking at 210 ms), and divided (peaking
at 195 ms). While there appeared to be qualitatively differences
in the strength and extent of the prediction effects across
Attention conditions, the interaction between Attention and
Prediction did not survive correction for multiple comparisons.
We then used a multiple sparse priors source reconstruc-
tion method to investigate the cortical regions that generated
the effects at the scalp level. Statistical parametric maps for
 1776 | Cerebral Cortex, 2018, Vol. 28, No. 5

Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020

Figure 3. Event-related potentials extracted from electrode Fz for each condition (mean/SEM). (A) The ERPs for each of the experimental conditions were extracted
from electrode Fz and compared over time. The grey shadings indicate the temporal widows during which a significant main effect of attention was found (**corrected
for the whole epoch, *corrected within the a priori MMN time window). (B) ERPs for attended and unattended prediction errors (the MMNs; i.e., the difference between
unpredicted and predicted) are plotted at electrode Fz. Grey shading indicates the temporal window during which a significant Attention by Prediction interaction
was found (*corrected within the a priori MMN time window).

source-reconstructed images revealed 2 significant clusters for
the main effect of Prediction in the left ([−42 −10 −38], peak-
level Tmax = 4.14, cluster-level PFWE = 0.019) and right inferior
temporal gyri ([44 0 −42], peak-level Tmax = 3.77, cluster-level
PFWE = 0.023) (Fig. 4C).
Opposition Wins Over Interaction—Evidence From
Posterior Probability Maps
Scalp level
BMS was used to compare the 2 competing models of the rela-
tionship between Attention and Prediction (the Opposition or
Interaction models; see Fig. 1). Specifically, we were interested
in comparing the strength of neural activation under the differ-
ent manipulations of attention and prediction. We used ran-
dom effects BMS to create group-level PPMs for each model,
derived from the log-model evidence of each participant, that
is, the evidence that a given model (Opposition or Interaction)
generated the data.
As shown in Figure 5, BMS revealed that the Opposition
model (“Attention and Prediction oppose”) was the more likely
(>75% model probability) explanation for the data across most
frontocentral channel locations at the majority of time points
(70–210 and 290–375 ms). However, the Interaction model
(“Attention and Prediction interact”) had a higher probability
(>75%) of explaining the data between 170 and 230 ms (i.e.,
within the MMN time window) at central and lateral parietal
channel locations. Thus, the relationship between Attention
and Prediction differed depending on both the time point and
scalp location; although more often than not, Attention and
Prediction had opposing effects.
The fact that the Interaction model won within the MMN
window and yet we did not find a significant interaction in
the classic GLM analysis could perhaps be explained by a
Prediction by Attention interaction effect that did not survive
 Modeling Attention and Prediction Garrido et al. | 1777

Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020

Figure 4. Main and simple effects of prediction at the scalp and source levels. (A) Spatiotemporal statistical analysis revealed significant effects of prediction (pre-
dicted vs. unpredicted) over frontocentral areas around 220 ms and over posterior parietal areas at 295 ms (displayed at P < 0.05, FWE whole-volume corrected). (B)
The effects of prediction across the 3 attentional manipulations revealed a prediction effect in the attended condition at 185 ms, the divided attention condition at
195 ms, and in the unattended condition at 210 ms, all located frontocentrally (displayed at P < 0.05, FWE whole-volume corrected). There was no significant interac-
tion (difference in the MMN between the attention conditions). (C) Source reconstruction analysis revealed a main effect of prediction within the left and right inferior
temporal gyri. (Displayed at P < 0.001 uncorrected and FWE corrected at the cluster-level.)

correction for multiple corrections. We further examined a
potential interaction effect, hindered perhaps by a rather con-
servative multiple comparison correction procedure. Firstly, we
used more lenient, uncorrected peak-level statistics to select 2
small interaction clusters at 175 ms (peak-level Fmax = 5.79,
peak-level Puncorr = 0.004; at central channels) and 360 ms (peak-
level Fmax = 5.45, peak-level Puncorr = 0.006; at right parietal
channels—see Fig. 6). We then took the spatiotemporal coordi-
nates of these clusters and extracted the posterior probability
of each model at that particular location. We constructed a 103
cube around these coordinates and took the average posterior
probability of each model over that volume. Our reasoning was
that if an interaction between Attention and Prediction were
present in the data, then the Interaction model would have a
higher posterior probability compared with the Opposition
model at these coordinates. We found that at 175 ms over fron-
tocentral channels there was a negligible difference between
the Opposition and Interaction models, with 48% and 52%,
respectively (Fig. 6). However, at 360 ms over the right lateral
parietal area, the Opposition model probability far exceeded
that of the Interaction model, with a value of 80%. Thus,
Attention and Prediction appear to have opposing effects later
in time.
Source Level
Finally, we applied the same BMS technique employed at the
sensor level to our source reconstructed results. BMS revealed
that the Opposition model had the higher model probability
and larger clusters at the source (Fig. 7). The Opposition model
achieved >50% model probability in the left middle temporal
gyrus (cluster size; KE = 82) and right inferior temporal gyrus
(cluster size; KE = 288). Conversely, the Interaction model
achieved > 50% model probability in a smaller cluster in the left
middle temporal gyrus (cluster size; KE = 32). We then com-
pared the model probabilities at the center of these clusters
and showed that the Opposition model was more probable
than the Interaction model in the left middle temporal and
right inferior temporal gyri (winning with 82% and 78% proba-
bility, respectively). Furthermore, model probabilities extracted
 1778 | Cerebral Cortex, 2018, Vol. 28, No. 5
Figure 5. Scalp Posterior Probability Maps of the Opposition and Interaction models over space and time. Maps display the posterior probability for both models, thre-
sholded at probability >75% over space and time. Scalp maps show the 4 time points with the largest significant clusters. The Opposition model wins (Attention and
Prediction oppose) across most frontocentral channels at the majority of time points (70–210 and 290–375 ms). The Interaction model wins (Attention and Prediction
interact) at the frontocentral and lateral parietal regions of the scalp (channel locations) between 170 and 230 ms.
Figure 6. Bayesian Model Comparison within the spatiotemporal clusters
extracted from the Prediction by Attention interaction. We extracted model
probabilities using the coordinates (scalp location and time points) of 2 clusters
from the Interaction results (based on the liberal threshold of P < 0.001 uncor-
rected). If interaction effects were present, the Interaction model would be
more likely to win over the Opposition model at these coordinates. At 175 ms
(within the MMN time window) and over central electrodes, there was a very
slight advantage for the Interaction over the Opposition model. However, at
360 ms over right lateral parietal channels, the Opposition model probability far
exceeded that of the Interaction model, with a probability of 80%.
from the peak of the Interaction model cluster showed only a
slight advantage for the Interaction over the Opposition model
(with 57% probability for the Interaction model) in the left mid-
dle temporal gyrus. Such a small difference between the proba-
bility of the Interaction model over the Opposition model at
this cluster suggests we should be cautious in drawing any
strong conclusions about its functional anatomy.
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
Dynamic Causal Modeling
The prior location of the cortical sources included in our DCMs
was based on MSP source reconstruction of ERPs corresponding
to the 4 conditions (attended standards, attended deviants,
unattended standards, and unattended deviants) of the Overall
Mismatch. Statistical parametric maps were inspected at a
more liberal threshold of P < 0.05 (uncorrected) to identify can-
didate neural sources of the effects observed on ERP amplitude
for the DCM analysis (Auksztulewicz and Friston 2015).
Following the selection of candidate sources, model structure
was optimized by comparing 2 alternative connectivity models
using data from each of the experimental conditions, with or
without bilateral connections between the left inferior tempo-
ral and inferior frontal gyri, with no between trial effects pres-
ent. Results indicated that the best model included recurrent
connection amongst all regions, that is, inputs to LA1 and RA1,
with LA1 connected to LITG, and LITG connected to LIFG, as
well as connections linking RA1 and RITG, and lateral connec-
tions between LITG and RITG. The selected model was then
used to further optimize condition-specific changes in the
extrinsic connectivity by comparing the types of extrinsic con-
nections present. Next, 15 competing models were tested
(Fig. 8A, B), each with a different subset of condition-specific
modulations of connections, according to the Opposition and
Interaction models on forward (F), backward (B), and recurrent
(R) connections (with, i, and without intrinsic modulations of
A1), as well as a null model precluding any modulations (N).
These models were fitted to each participant’s data to explain
observed differences in ERP amplitude. Random-effects BMS
revealed that the Opposition model with modulation of forward
connections outperformed all other models (Fig. 8C).
Discussion
In this study, we adjudicated between 2 alternative computa-
tional models of the effect that spatial attention has on expec-
tations. Using Bayesian model comparison of scalp PPMs we
found that, except for an early time window (within the typical
MMN), the Opposition model won over the Interaction model.
This suggests that, for the most part, attention provides an
equivalent boost to neuronal responses to predicted and
 Modeling Attention and Prediction Garrido et al. | 1779

Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020

Figure 7. Source posterior probability maps of the Opposition and Interaction models (top) and model probabilities for the 3 major clusters of the 2 models (bottom).
BMS was used for model inference at the group-level for the source reconstructed images. Here, the Opposition model achieved >50% model probability in the left
middle temporal (cluster size; KE = 82) and right inferior temporal (cluster size; KE = 288) gyri. The Interaction model achieved >50% in a small cluster in the left mid-
dle temporal gyrus (cluster size; KE = 32). Overall, the Opposition model achieved higher probability over a larger number of voxels. Extraction of model probabilities
from the peak of the Opposition clusters showed that this model won with 82% probability in the left middle temporal gyrus and with 78% probability in the right
inferior temporal gyrus. Model probabilities extracted from the peak of the Interaction cluster showed minimal differences between either model at this location,
with 57% posterior probability for the Interaction model. Note the differences in the color map scales between the Opposition and Interaction models.

unpredicted stimuli. Similarly, at the source level we found
stronger evidence for the Opposition model underlying a fron-
totemporal network. We investigated this further with DCMs
that employed trial-dependent plastic changes according to
either the Opposition or the Interaction model. In agreement
with the model-based scalp and source analysis, we found that
the family of Opposition models better explained the data.
Classic SPM analysis of spatiotemporal maps revealed an effect
of prediction across and within all attentional manipulations,
which peaked within the typical MMN time window and at
frontocentral channels. This effect was statistically greater in
the attended compared with the unattended conditions at the
single channel level, where MMN is typically seen, suggesting
that attention amplifies prediction errors. At the whole spatio-
temporal map level, however, this interaction effect did not
survive correction for multiple comparisons over the whole
space-time, despite the appearance of somewhat larger clusters
for the attended than the unattended condition,
Our finding of a prediction error effect in all attention condi-
tions (attended, unattended, and divided) is in agreement with a
vast body of work suggesting that the MMN is elicited regardless
of attention, and hence is “pre-attentive” in nature (Naatanen
et al. 2001). This is in contradistinction to Auksztulewicz and
Friston (2015), who did not find an effect of prediction in the
absence of attention (although this might have been due to a
lack of power, as very few trials were included). Again, our
finding of a prediction error effect regardless of attention is oppo-
site to Todorovic et al. (2015), who found that while beta syn-
chrony decreased with expectation in the unattended condition,
no difference was found in the attended condition. The latter is
seemingly at odds with the idea that attention amplifies predic-
tion errors as previously shown (Jiang et al. 2013;
Auksztulewicz and Friston 2015), and as revealed in the current
study. A number of factors could explain such conflicting
results. Perhaps most importantly, very different paradigms
and measures were employed across the relevant experiments.
Both our study and that of Auksztulewicz and Friston (2015)
investigated the effects of attention and prediction on evoked
responses in an oddball paradigm, whereas Todorovic et al.
(2015) focused on endogenous oscillatory activity. Moreover,
both Auksztulewicz and Friston (2015) and Todorovic et al.
(2015) manipulated temporal attention, whereas here we
manipulated spatial attention. Finally, in our experiment atten-
tion and prediction were manipulated within the same spatial
location (left or right ears), but were drawn toward independent
auditory “objects” (noise for the attention task, and tones for
the concurrent oddball stream). By contrast, the aforemen-
tioned studies (and that of Kok et al. (2012)) manipulated atten-
tion and prediction within the same (visual or auditory) object. It
is possible that our attention manipulation, based on spatial
selectivity, had a small effect on the tones (in the attended con-
dition), given that these were task-irrelevant and that they
 1780 | Cerebral Cortex, 2018, Vol. 28, No. 5

Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020

Figure 8. Dynamic Causal Modeling hypotheses testing for plastic changes according to the Opposition and Interaction families of models. (A) Eight model architec-
tures were considered, which tested for trial-specific modulation in forward (F), backward (B), and recurrent (R, i.e., both forward and backward) connections, as well
as a null model precluding any modulations (N) in extrinsic connections. These models were considered with (bottom row) and without (top row) intrinsic modula-
tions of A1 (subscript i). The nodes in the model included bilateral primary auditory cortices (LA1 and RA1), bilateral inferior temporal gyri (LITG and RITG) and left
inferior frontal gyrus (LIFG). (B) Extrinsic connectivity was optimized using responses in all 4 conditions under the Opposition or Interaction model. A total of 15 com-
peting models were tested, each with a different subset of condition-specific modulation of connections, according to the Opposition (blue) and Interaction (green)
models on F, Fi, B, Bi, R, Ri, and Ni. N did not include any modulations (orange). Summed model exceedance probabilities across each family show the winning family
as the Opposition Family (left; blue). (C) The winning model architecture had recurrent connections between all regions, intrinsic modulation of A1, lateral connec-
tions between bilateral ITG, and included condition-dependent modulations according to the Opposition model in the forward connections (blue lines).

never occurred at the same time as the task-relevant noise
gaps. However, we believe that this is improbable for 2 reasons.
First, the onset of the noise gaps was unpredictable and hence
participants had to constantly monitor the stream of sounds on
the task-relevant side of space. Second, it is unlikely that partici-
pants learned that the noise gaps never coincided with the tones,
and could therefore momentarily disengage attention from the
noise task. Having said that, the possibility remains that by hav-
ing the participants focus on the noise streams instead of the
tones, our attention manipulation might not have influenced the
neural representations of the tones as much as it would have,
had we asked the participants to focus on the tones. Future work
should test whether manipulating attention and prediction for
common versus independent stimuli alters the extent to which
they interact.
In this work we directly compared 2 competing models of
the effects of attention on expectations—the Interaction and
Opposition models—put forward in Kok et al. (2012). The data
in that study were consistent with the Interaction model when
considering regions of the visual cortex (V1, V2, and V3). Here,
however, we took a different approach by implementing the
models computationally and directly testing them against our
data. By using Bayesian model comparison of statistical maps
of EEG activity, and DCMs for ERPs, we were able to quantify
how likely each of these 2 models was at every point of space
and time at the scalp level, at each voxel in source space, and
in the trial-dependent plastic changes within a cortical net-
work. The Opposition model was unambiguously favored in
our data at every level, that is, scalp, source, and network. At
the network level we found that the plastic changes according
to the Opposition model were more pronounced in forward
connections. This is consistent with the idea that attention
boosts, or heavily weights, prediction errors, which are then
conveyed upward in the cortical hierarchy. Such prediction
errors signal the need to update an internal perceptual model
of the world, in turn prompting learning. At first glance it may
appear that boosting of prediction errors is more consistent
with the Interaction model. It is important to note, however,

that the corollary of the Interaction model is that attention
reverses prediction, such that larger responses will be observed
for predicted compared with unpredicted stimuli. In this sense,
attention changes the sign of the prediction error instead of
boosting it. On the contrary, boosting of prediction errors could
in principle be accommodated by the Opposition model as it
predicts a larger difference between unpredicted and predicted
responses in the attended versus unattended condition. Having
said this, our instantiation of the Opposition model is agnostic
to such a relationship and was not modeled explicitly here. The
Opposition model simply assumes that unpredicted responses
will always be larger than predicted responses, regardless of
attention, and that attention will boost these responses. It may
also appear surprising that our attention manipulation did not
modulate backward connections, which are thought to convey
updated predictions (Friston 2005). In our paradigm, however,
the predictions did not require constant updating, unlike in
some other paradigms in which the rule constantly changes,
such as in roving MMN (Garrido et al. 2008) or in reversal learn-
ing (Ghahremani et al. 2010). In such scenarios, it is possible
that attention would modulate prediction updating via feed-
back processing (Desimone and Duncan 1995; Spratling 2008).
We should, however, be cautious when interpreting the find-
ings from our best individual model. While we have good evi-
dence for an advantage of the Opposition family (77%) over the
Interaction family (17%), and thus can assert that attention and
prediction have opposing effects on plastic changes, we are
less confident about where exactly in the network these effects
might be expressed, given the relatively small advantage for
the forward model over the remaining models tested.
While the better performance of the Opposition over the
Interaction model is generally at odds with the findings by Hsu
et al. (2014) and Kok et al. (2012), there was a narrow window of
agreement in which the Interaction model was better at
explaining the data at the scalp level, perhaps tellingly within
the MMN time frame. This is an interesting finding, as it seems
to point to a tonic Opposition effect between Attention and
Prediction, and a phasic Interaction effect. Again, there are differ-
ences in both the type of paradigm and the neuronal measures
between our study, which used EEG, and the experiment of Kok
et al. (2012), which used fMRI. Although Attention was manipu-
lated spatially in both studies, in our study it was directed
towards a different (instead of the same) object. Moreover, our
Prediction manipulation was learnt from the sequence of sti-
muli, rather than instructed (as in Kok et al. (2012)).
In conclusion, our findings provide empirical evidence for a
computational model of the opposing interplay of attention
and expectations in the brain. These opposing effects are mani-
fested in neuronal activity and in plastic changes within a fron-
totemporal network engaged in sensory prediction errors. We
demonstrate that attention boosts neuronal responses to pre-
dicted and unpredicted stimuli, and replicate the finding that
attention boosts prediction errors, in keeping with the predic-
tive coding framework (Rao and Ballard 1999; Friston 2005).
Finally, we demonstrate that prediction errors are elicited
regardless of one’s state of attention, providing further support
to the idea of a preattentive nature of change detection systems
in the brain (Naatanen et al. 2001).
Funding
Australian Research Council (ARC) Discovery Early Career
Researcher Award (DE130101393) and a University of Queensland
Fellowship (2016000071) to M.I.G., an ARC Australian Laureate
Modeling Attention and Prediction Garrido et al. | 1781
Fellowship (FL110100103) to J.B.M., the ARC Centre of Excellence
for Integrative Brain Function (ARC Centre Grant CE140100007) to
M.I.G. and J.B.M., and an ARC Special Research Initiative—Science
of Learning Research Centre (SR120300015) to J.B.M.
Notes
We thank the volunteers for participating in this study and
Maria Joao Rosa for discussions. Conflict of Interest: The authors
declare no competing financial interests.
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020
References
Auksztulewicz R, Friston K. 2015. Attentional enhancement of
auditory mismatch responses: a DCM/MEG study. Cereb
Cortex. 25:4273–4283.
Bekinschtein TA, Dehaene S, Rohaut B, Tadel F, Cohen L, Naccache
L. 2009. Neural signature of the conscious processing of audi-
tory regularities. Proc Natl Acad Sci USA. 106:1672–1677.
Desimone R, Duncan J. 1995. Neural mechanisms of selective
visual attention. Annu Rev Neurosci. 18:193–222.
Feldman H, Friston KJ. 2010. Attention, uncertainty, and free-
energy. Front Hum Neurosci. 4:215.
Friston K. 2005. A theory of cortical responses. Philos Trans R
Soc Lond B Biol Sci. 360:815–836.
Garrido MI, Friston KJ, Kiebel SJ, Stephan KE, Baldeweg T, Kilner
JM. 2008. The functional anatomy of the MMN: a DCM study
of the roving paradigm. Neuroimage. 42:936–944.
Garrido MI, Kilner JM, Kiebel SJ, Stephan KE, Friston KJ. 2007.
Dynamic causal modelling of evoked potentials: a reproduc-
ibility study. Neuroimage. 36:571–580.
Garrido MI, Sahani M, Dolan RJ. 2013. Outlier responses reflect
sensitivity to statistical structure in the human brain. PLoS
Comput Biol. 9:e1002999.
Ghahremani DG, Monterosso J, Jentsch JD, Bilder RM, Poldrack
RA. 2010. Neural components underlying behavioral flexibil-
ity in human reversal learning. Cereb Cortex. 20:1843–1852.
Henson RN, Mattout J, Phillips C, Friston KJ. 2009. Selecting for-
ward models for MEG source-reconstruction using model-
evidence. Neuroimage. 46:168–176.
Hsu YF, Hamalainen JA, Waszak F. 2014. Both attention and
prediction are necessary for adaptive neuronal tuning in
sensory processing. Front Hum Neurosci. 8:152.
Jiang J, Summerfield C, Egner T. 2013. Attention sharpens the
distinction between expected and unexpected percepts in
the visual brain. J Neurosci. 33:18438–18447.
Kok P, Rahnev D, Jehee JF, Lau HC, de Lange FP. 2012. Attention
reverses the effect of prediction in silencing sensory signals.
Cereb Cortex. 22:2197–2206.
Mattout J, Henson RN, Friston KJ. 2007. Canonical source recon-
struction for MEG. Comput Intell Neurosci. 2007: Article ID
67613.
Montague PR. 1999. Reinforcement learning: an introduction.
Trends Cogn Sci. 3:360–360.
Naatanen R, Alho K. 1997. Higher-order processes in auditory-
change detection. Trends Cogn Sci. 1:44–45.
Naatanen R, Tervaniemi M, Sussman E, Paavilainen P, Winkler
I. 2001. “Primitive intelligence” in the auditory cortex.
Trends Neurosci. 24:283–288.
Opitz B, Mecklinger A, Friederici AD, von Cramon DY. 1999. The
functional neuroanatomy of novelty processing: integrating
ERP and fMRI results. Cereb Cortex. 9:379–391.
Penny WD, Trujillo-Barreto NJ, Friston KJ. 2005. Bayesian fMRI time
series analysis with spatial priors. Neuroimage. 24:350–362.
 1782 | Cerebral Cortex, 2018, Vol. 28, No. 5
Rao RP, Ballard DH. 1999. Predictive coding in the visual cortex:
a functional interpretation of some extra-classical recep-
tive-field effects. Nat Neurosci. 2:79–87.
Rosa MJ, Bestmann S, Harrison L, Penny W. 2010. Bayesian
model selection maps for group studies. Neuroimage. 49:
217–224.
Spratling MW. 2008. Reconciling predictive coding and biased
competition models of cortical function. Front Comput
Neurosci. 2:4.
Summerfield C, Egner T. 2009. Expectation (and attention) in
visual cognition. Trends Cogn Sci. 13:403–409.
Summerfield C, Koechlin E. 2008. A neural representation of
prior information during perceptual inference. Neuron. 59:
336–347.
Todorovic A, Schoffelen JM, van Ede F, Maris E, de Lange FP.
2015. Temporal expectation and attention jointly modulate
auditory oscillatory activity in the beta band. PloS One. 10:
e0120288.
Downloaded from https://academic.oup.com/cercor/article-abstract/28/5/1771/3571164 by Biology Library user on 03 March 2020